ESTRO 2024 - Abstract Book

S3134

Physics - Autosegmentation

ESTRO 2024

Figure 1. An example of (A) input CT and ce-CT, (B) preprocessed images, (C) result of the GTVs model, and (D) result of the OARs model.

Results:

Table 1 shows a summary of the achieved results over the training, validation, and test phases for each of the two tasks in terms of the Dice (DSC) and normalized surface Dice (NSD) metrics.

Table 1. The overall quantified metrics (mean) over different subsets.

In detail, out of 54 structures, 21 OARs were segmented with a mean DSC larger than 0.90, 23 structures were segmented with a mean DSC between 0.80 to 0.90, and 10 structures ended up with a mean Dice less than 0.80 from which only the left and right hippocampus as well as the right middle ear bone were segmented poorly (less than 0.60). It is worth mentioning that the computational time in the inference phase to fully segment all the described structures takes less than two minutes.

Conclusion:

While manual delineation of 45 OARs and 2 GTVs of H&N cancer is prone to inconsistencies and potential errors, labor-intensive, and time-consuming, we proposed an accurate and robust model for automatic segmentation of these structures. Our model outperforms state-of-the-art methods by winning the SegRap 2023 challenge and yields accurate and robust segmentation results for both GTVs and OARs structures.

Keywords: head&neck, organ segmentation, tumor segmentation

References: